Sorry, I missed this. I have already updated the docs.

The reason why it is not encouraged is because it tells all loggers to stop what they are doing and write to disk. If you call it several times, they will stop and write to disk, every single time, even though they believe batching and waiting is the efficient thing to do.

Of course, we can always improve the docs and PRs are welcome. If this is happening quite a bit, then we are not being told about those cases, because they would certainly have been addressed otherwise.

Are you sure this is what’s happening?

I tried crashing a process by raising an exception right after logging and it works as expected:

iex> require Logger

iex> spawn(fn ->
...>   Logger.info("info")
...>   raise "oops"
...> end)

00:58:27.917 [info] info
00:58:27.919 [error] Process #PID<0.109.0> raised an exception
** (RuntimeError) oops

Maybe your system is getting overloaded?

Also looking into the actual implementation of Logger.flush/0 it seems to be no-op by default: elixir/lib/logger/lib/logger.ex at 78f63d08313677a680868685701ae79a2459dcc1 · elixir-lang/elixir · GitHub – because the default logging destination for logger_std_h (which Logger uses) is starndard_io and not a file – but I haven’t actually verified it, just took a look at the code